County-level corn yield prediction using deep transfer learning
Abstract
Corn is one of the largest commodities grown by land area in the United States and timely and reliable estimation of corn yield is imperative for ecology, economics and human society. The publicly available predictions are mainly at country or state level, which are derived based on historical statistical relationships and in-season surveys. To further improve the utility of crop yield prediction models, it is necessary to establish them at a finer spatial scale, such as at the county level. Recently, different machine learning models have been developed using remote sensing observations for yield prediction. Though successful, a main drawback of the existing approaches is that the models established between reported yields and remote sensing measurements are location-specific and have limited applicability to other regions where do not have extensive ground truth data. To address this issue, in our study, a deep transfer learning model based on Bayesian Neural Networks (BNN) will be developed to improve the model transferability across different spatial domains. The proposed model will not only provide the predicted yields, but also can indicate the uncertainty of the prediction results in the new spatial domain. Specifically, the yield records, which are considered as the ground truth data, will be collected from the USDA NASS at a county level from 2001 to 2018. In addition, a large set of potential yield determinants derived from three data sources will be used as input for the model development, including 1) MODIS 16-day 250m Enhanced Vegetation Index (EVI) and 8-day 1-km Land Surface Temperature (LST); 2) DAYMET gridded climate variables (e.g. temperature and precipitation) at 1-km spatial resolution; 3) Soil properties (e.g. soil depth, soil texture, organic carbon content) based on SoilGrid data. For evaluating the model transferability, the counties understudy will be divided into two diverse ecological regions based on the United States Environmental Protection Agency (EPA), including the 1) Eastern Temperate Forests region, which is mainly covered by dense and diverse forests, and 2) Great Plain region, which mainly consists of flat grasslands and has a scarcity of forests. The experiments will be carried out by training the model in one region and testing on the other.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.B54D..02M
- Keywords:
-
- 0402 Agricultural systems;
- BIOGEOSCIENCES;
- 0434 Data sets;
- BIOGEOSCIENCES;
- 0480 Remote sensing;
- BIOGEOSCIENCES;
- 1842 Irrigation;
- HYDROLOGY